Deep and superficial SEEG shaft halves had similar key-press AUC, but both scored below all pairs

In archived features from 8 subjects in OpenNeuro ds004473 v1.0.2, we compared offline key-press-versus-rest detection using deep (tip-side) and superficial (entry-side) bipolar pairs from stereoelectroencephalography (SEEG) shafts. We used contact number as a depth proxy, split each eligible shaft into equal-count halves, and trained Ledoit-Wolf shrinkage linear discriminant analysis (LDA) on log mean 70–150 Hz high-gamma (HG) and 13–30 Hz beta power. Training used trials 1–100; testing used trials 101 onward (32–198 test trials per subject). The primary mean test-AUC difference (superficial minus deep) was −0.004 (4/8 subjects positive; exact two-sided sign-flip p = 0.81; subject-bootstrap 95% CI [−0.036, 0.023]). Superficial and deep halves each scored below all pairs: superficial minus all was −0.038 (0/8 positive; p = 0.008; 95% CI [−0.059, −0.019]) and deep minus all was −0.034 (1/8 positive; p = 0.016; 95% CI [−0.047, −0.019]). The superficial half did not meet the planned 0.02 AUC non-inferiority margin against all pairs. With HG only, superficial minus deep was −0.034 (4/8 positive; p = 0.32; 95% CI [−0.091, 0.022]), and superficial minus all was −0.071 (1/8 positive; p = 0.039; 95% CI [−0.121, −0.022]). Single-pair HG AUC across five normalized shaft-position bins ranged from 0.525 at the tip to 0.536 in the middle bin and 0.534 near entry (831 pairs). These offline results in 8 clinical epilepsy subjects do not show a difference between the two shaft halves. The half-versus-all loss is confounded by the smaller number of pairs; equal-size random halves were not tested. Contact number is not anatomy, few pairs were near motor cortex, and subject 11 had only 32 test trials.

stereoelectroencephalography · brain-computer interface · key-press detection · linear discriminant analysis · high-gamma power · depth · AUC